000			04596ntm a2200337 i 4500
003			MY-KuUP
005			20260403120718.0
006			t||||fr|||| 000 0
007			ta
008			260403t20252025my a|||fr|||| 000 0 eng d
020			_aTHE0010429 (Local) _qHardback
040			_aUMPSA _beng _cUMPSA _erda
090			_aFTKPM .J47 2025 r Thesis
100	0		_aJessnor Arif Bin Mat Jizat, _eauthor.
245	1	0	_aClassification of wafer defect images by synthesizing transfer learning models and machine learning classifiers using different input transformations / _cJessnor Arif Bin Mat Jizat
264		1	_aKuantan, Pahang : _bUMPSA , _c2025
264		4	_c© 2025
300			_axv, 145 pages : _billustrations ; _c30 cm. + _e1 CD-ROM
336			_2rdacontent _atext
337			_2rdamedia _aunmediated
338			_2rdacarrier _avolume
347			_2rda _atext file _bPDF
500			_aFaculty of Manufacturing & Mechatronics Engineering Technology
502			_aThesis (Doctor of Philosophy) -- Universiti Malaysia Pahang - 2025
504			_aIncludes bibliographical references
520	3		_aSilicon wafers are a critical material in the semiconductor industry, making wafer defects a significant concern for production and company revenue. Traditionally, defect inspection has been a manual process where personnel sample each wafer batch. This labour-intensive method often leads to inconsistencies between inspectors and is prone to errors from human fatigue. As a continuous improvement initiative, automated optical inspection machines are now being deployed. These systems typically use image processing to differentiate between defective and non-defective wafers. However, since wafer defects are often miniscule, they can be missed by conventional techniques. This thesis explores the integration of a deep learning algorithm to improve defect identification. The study assesses the performance of Transfer Learning models and Machine Learning classifiers for wafer defect classification. It evaluates their robustness across different input image transformations and against image compression. A dataset of 500 images was used, featuring four defect categories—Burnt Mark, Missing, Scratch, and Contamination—along with one non-defective category. The dataset was split into a 60:20:20 ratio for training, validation, and testing. The experiment evaluated 14 Transfer Learning models—DenseNet121, DenseNet169, DenseNet201, InceptionResNetV2, InceptionV3, MobileNetV2, NasNetLarge, NasNetMobile, ResNet101V2, ResNet152V2, ResNet50V2, VGG16, VGG19, and Xception—alongside four Machine Learning classifiers: Support Vector Machine (SVM), Random Forest, Logistic Regression, and k-Nearest Neighbour (k-NN). The image transformations applied were Greyscale, Histogram Equalisation, Convolution Filter, and Local Binary Pattern, while Fast Fourier Transformation (FFT) was used for image compression. Model performance was evaluated based on Accuracy, Precision, Recall, F1-Scores, and training time. The experimental results yielded several key findings: with untransformed images, the InceptionV3-Logistic Regression pipeline delivered the best performance at 95% accuracy, 95.2% average precision, and 95.0% average recall, closely followed by the NasNetLarge-SVM model which achieved comparable accuracy and recall with a faster training time of 273.01 ms. For transformed images, the highest performance was attained using convolution filters with the InceptionResNetV2-Logistic Regression model, yielding 95% accuracy and 95.6% average precision. Furthermore, in image compression tests, the InceptionV3-Logistic Regression pipeline demonstrated remarkable robustness, maintaining 94% accuracy even with a 95% compression rate via FFT, simultaneously reducing storage requirements from 99.4MB to 38.4MB. In conclusion, this thesis successfully demonstrates that the proposed classification pipelines maintain robust performance under significant image compression and various input alterations. The finding that classification accuracy remains above 90% even with 95% FFT compression is a crucial advancement for practical implementation. This directly addresses the critical challenge of storage efficiency, enabling the reliable deployment of transfer learning and machine learning pipelines in industrial settings where large image collections are necessary.
610	2	0	_aFaculty of Manufacturing & Mechatronics Engineering Technology _xDissertations
650		0	_aUniversities and colleges _xDissertations
650		0	_aTheses _xDissertations
942			_2lcc _cTHESIS
999			_c104570 _d104576